A clamp is a limit, not a gain — the [occupancy] table

core/tests/saturation.rs is a HARD gate: it walks every shipped preset’s expressions over 12 s of dynamic:110 and fails the build on any clamp() whose inner value sits at its upper bound for 90 % or more of the hops (ADR-0062).

That number is called occupancy, and it names the one defect nothing else here can see. Write

warp = "clamp(bass * 16, 0, 0.3)"

and the ceiling is reached at bass = 0.019 — below the quietest hop of any real material. The binding is not a parameter any more; it is the constant 0.3 with a brief flicker near silence. Every reactivity instrument in this project scores it as perfectly reactive, because they all compare a driven band against silence, and against silence a binary switch is maximally responsive. What that costs has been measured across the whole library at once: 263 of 332 clamped band terms in exactly this state, 14 presets with no live audio term at all, behind a fully green suite.

Driving the gate with real audio does not change this. reactivity puts real PCM through the real analyzer rather than hand-built frames — but it still compares a driven band against silence, and that is the property this section is about. A saturating clamp is maximally reactive against silence whether the “driven” number came from an FFT or from a literal. saturation is the only gate that asks the question this section asks.

The fix is always the same arithmetic: divide the gain until the bound is reached only on peaks. With bass averaging 0.661, clamp(bass * 0.45, 0, 0.3) reaches the ceiling around bass = 0.67 — on hits, and nowhere else. The failure message states the level the ceiling is already reached at, so you can do the division from the failure alone.

Check before you commit — occ is the column, and there is a SAT line per finding:

cargo run -p standalone --example shot -- --report --presets presets

The healthy library sits well under the threshold: across 339 clamped bindings the highest occupancy is 0.61 and the next is 0.44.

If the pin is the design, say so in the preset. A safety rail that exists to bind at peak is not this defect:

[occupancy]
exempt = ["fade"]   # this clamp is a rail, not a gain: pinning is intended

An exemption silences the gate and nothing else — the binding still appears in --report’s occ count and its SAT line, so it stays visible in review. Reach for it when you can say in one sentence why the bound must hold; reach for the division otherwise. A name no [params] entry binds is inert and warns at load.